KL School of AI 15th December 2018 The promise of Generative Adversarial Networks (GANs)

1. THE AGE OF GANs
KL SCHOOL OF AI PRESENTS
By: Faris Hassan - farisology
Generative Adversarial Networks

2. Farisology
I'm Faris Hassan
Graduated from MMU majored in AI
Research in medical imaging and
automated
diagnoses systems for Tuberculosis using
chest radiographs
Worked in Sedania Innovator as a data
Scientist.
Now joined Fave.

3. Agenda
Topics to be Discussed
GANs and Fake stuff
Applications of GANs
GANs Concept
The Arch. of GANs.
The difference with Neural Nets.
hands-on a bit
New Archs. and variants of GANs.
Disadvantages of GANs.
Steps of build Generative models.
Tips for better performance

4. Have you heard of Deep Fake?
You can say it one of the creepiest things you will know about AI. If you think that
detecting fake news have is a headache, you will be freaked out by deep fake. It is
fake in totally new level.
The fancy definition: Deepfake, a portmanteau of "deep learning" and "fake", is an
artificial intelligence-based human image synthesis technique. It is used to combine
and superimpose existing images and videos onto source images or videos. It is an
artificial intelligence assisted technology that make the creation of fake videos very
convincing. Watch the video: Deepfake

5. Are you scared now ?
Why am I sharing this ?

6. Democratizing
Artificial
Intelligence
Will bring more good and transform humanity
BECAUSE

7. Generative Adversarial
Networks
The idea of adversarial
networks have been around for
awhile but there hasn't been
any development or a
significant development till a
research published in 2014 by
the scholar of OpenAI Ian
Goodfelow.

10. GANs
Concept
Unlike a conventional neural networks GANs
have a game theoretic approach. The
structure involves the training of two neural
networks. The networks learn from a training
distribution via a 2 player game. The two
components are the generator and
discriminator.

11. These two gonna be in a constant battle through out the training process till the
forger make awesome fake stuff

12. Two competing
networks
The Struggle is real bro...
There are two networks competing in this
unsupervised learning environment. The aim of the
generator is to create some representation of input
using a random seed - usually Gaussian noise - and
present it to the discriminator. The discriminator
then decide which of the generated data is real or
fake.

13. The nash equilibrium in
this scenario is realized
when the generator
learns enough to make
the discriminator
unable to recognize the
fake data.

14. GAN & NEURAL NET
The
difference
with neural
networks.
The major difference lies in the need of cost
function. Traditionally cost function has to be
carefully designed by human engineers. But
adversarial networks learn their own cost
function (based on each other’s feedback).
Discriminator network tries to learn the
boundary between the classes so that it can flag
the fake data. Generator network tries to learn
the distribution of class.

15. WHEN YOUR DATA IS
NOT ENOUGH ....
GANs are
too much till
you need a
GAN.
Deep learning applications require huge amount
of data for training. The availability of huge
training dataset has always been the pain point
not as easy task for people in several fields. GANs
can come to rescue in this problem. We can use
GANs to generate our own datasets using little
examples.
The quality might not be very accurate but
people have started employing GANs for this
task.

16. In Code we Trust
Let's build a
standard GAN
Hands-on

17. There are
more GANs
than Gangs
just saying
GAN Archive
Press here if you wish

19. Reality hits hard

20. 3 MAJOR PROBLEMS WITH GAN

21. 3 MAJOR PROBLEMS WITH GAN

22. 3 MAJOR PROBLEMS WITH GAN

23. DISADVANTAGES OF GAN
issues
Unstable to train
you have two networks
using the same
backpropagation
Cannot perform
inference queries or
seek any form of
interpretation
it requires a clear
objective, otherwise
you won't know what is
it doing

24. Steps to build cool GAN
Step1: Define the problem.
Step2: Define architecture.
Step3: Train Discriminator on real data for n epochs.
Step4: Generate fake inputs for generator and train discriminator on
fake data.
Step5: Train generator with the output of discriminator.
Do Step3 - Step5 for n epochs.
Step6: Check the fake data manually. Do they look legit? If not get back
to Step3-Step5 for more epochs till you have some awesome fake
stuff.

25. How to make GANs
work?
Tips
1- Normalize your input.
2- Use DCGAN when you can because its
always better than just GAN.
3- Use SGD for discriminator and ADAM for
generator.
4- Read more from GAN-Hacks

26. Resources:
1- Get your hands on more DCGAN
2- Build your own celebrity GANs
3- You can also do it in pytorch much easier. PYTORCH

27. ASK ME QUES.....
Also download fave for awesome deals
farisology
Use this promo-code for 20%

28. Contact me
Github
github.com/farisology
Email Address
fareshasan.ai@gmail.com
Phone number
012 5xx or just telegram la